Early Continuous Electroencephalography, Clinical Parameters, and Short-Term Functional Outcomes in Pediatric Traumatic Brain Injury: Single-Center, Retrospective Cohort, 2010–2020

Abstract

OBJECTIVES:

Traumatic brain injury (TBI) is a leading cause of pediatric morbidity and mortality. This study first investigates clinical characteristics and continuous electroencephalography (cEEG) parameters associated with short-term functional outcomes in pediatric patients following TBI. Second, we use these data for a hypothesis-generating model about outcomes.

DESIGN:

Retrospective cohort study.

SETTING:

PICU within a quaternary care hospital.

PATIENTS:

Pediatric patients (< 18 yr) admitted from January 2010 to December 2020 with TBI who underwent cEEG within 72 hours of admission.

INTERVENTIONS:

None.

MEASUREMENTS AND MAIN RESULTS:

Patient demographics, clinical parameters, hospital course, and cEEG features were reviewed for associations with mortality and new disability (as defined by change in Functional Status Score between admission and discharge of ≥ 3 points). A statistical prediction model for patient outcome was created combining cEEG parameters with admission Glasgow Coma Scale (GCS) score and radiographic findings. We included 142 patients: 100 (70%) of whom had no new disability at hospital discharge; 42 (30%) had a new disability, including eight deaths. Univariate analysis showed favorable outcomes were associated with normal electroencephalogram background, reactivity, and sleep features (p < 0.001 for each). A model inclusive of these electroencephalogram parameters and GCS had high predictive ability for outcome with 0.94 with 95% CI (0.90–0.98).

CONCLUSIONS:

Specific cEEG findings observed acutely after injury, in combination with other clinical characteristics, may serve as biomarkers for short-term functional outcomes after pediatric TBI. Further validation of the model in another population is now required.

United States estimated of hospitalized children in 2006 (1) and country-specific global data from 1995 to 2015 (2) demonstrate that traumatic brain injury (TBI) is an important global public health problem that impacts more than 3 million children every year. However, prognostication of outcomes following pediatric TBI is challenging, especially given the heterogeneous mechanisms of injury and developmental plasticity of children. Furthermore, while studies have focused on mortality as the primary outcome for children with severe TBI (3-5), there is an increased appreciation for functional outcomes in survivors; morbidities, such as neurocognitive impairments, motor deficits, and social-emotional disturbances affect the child and immediate family and result in a large economic burden (6). Determining statistical predictors of short-term functional outcome could impact acute care for patients and improve family counseling.

Most prognostic models include readily available demographic factors and clinical metrics including age, sex, Glasgow Coma Scale (GCS) score, pupillary reaction, and radiographic findings (7-10). As continuous electroencephalography (cEEG) is increasingly used following acute brain injuries to surveil for seizures, other characteristics such as background continuity, reactivity, and sleep features may provide insight into cortical functioning (11-14). Indeed, cEEG metrics have served as biomarkers for outcome prediction following injuries such as pediatric cardiac arrest (15, 16), pediatric TBI, and adult TBI (17-20). Interestingly, a 2014–2020 cohort study in pediatric patients found associations between quantitative electroencephalogram metrics and both intracranial hypertension and functional outcomes following TBI (20), consistent with a predictive relationship between cortical function and injury severity. Here, we hypothesized that qualitative cEEG-derived metrics could be combined with other early, accessible clinical data (i.e., GCS, radiographic findings) to generate a prognostic model of functional outcomes following pediatric TBI.

METHODS

This retrospective study was considered by the institutional review board (IRB) at Children’s National Hospital and approved on February 7, 2021 (study title “Determining best practices in the neurologic care of critically ill children” with study number Pro00015473). Informed consent was not required. During this study, all procedures were followed in accordance with the ethical standards of the IRB and our institutional committee on human experimentation, and with the Helsinki Declaration of 1975.

The cohort studied included pediatric patients (< 18 yr) admitted to the Children’s National Hospital PICU with a diagnosis of TBI between January 2010 and December 2020. Children’s National Hospital is the primary freestanding level 1 Pediatric Trauma Center in the Washington, DC, metropolitan area. Patients with TBI were identified using International Classification of Diseases, 9th Edition and International Classification of Diseases, 10th Edition codes, which were then verified on chart review. We included in the analysis, those who had at least 8 hours of cEEG monitoring within 72 hours of admission. Patients were excluded if they had a history of a prior seizure disorder or if they suffered a cardiac arrest in association with their TBI, as both would impact electroencephalogram findings. Admission to the PICU was at the discretion of the clinical teams based on clinical presentation, injury severity, and the need for frequent neurologic assessments.

Demographic data, injury characteristics, neuroimaging results, cEEG features, neurosurgical intervention, PICU length of stay (calculated as discharge date minus admission date), and discharge disposition were collected. Study data was managed using Research Electronic Data Capture tools hosted at Children’ National Hospital. The GCS score was recorded from admission to the emergency department. When multiple GCS scores were documented, the score documented by the lead trauma surgeon was used. The admission and discharge Functional Status Scores (FSSs) were calculated by review of the medical record using methods previously published (21). The FSS score is a cumulative result of six domains (i.e., mental, sensory, communication, motor, feeding, and respiratory) that each has a score from 1 (no dysfunction) to 5 (severe dysfunction) with the sum FSS ranging from 6 to 30 (22). The FSS is closely associated with the Pediatric Cerebral Performance Category and Pediatric Overall Performance Category scales, and an increase of FSS greater than or equal to 3 points was defined as the development of a new disability during the admission (21-23).

At our institution, cEEG monitoring is applied for patients with acute brain injury and altered mental status according to the American Clinical Neurophysiology Society guidelines, with the final decision as to whether cEEG is performed being at the discretion of the clinical team caring for the patient (15). cEEG electrodes were placed according to the standard international ten of 20 system of placement and interpreted by board certified pediatric epileptologists and neurophysiologists in real-time and at set intervals for patient safety. The electroencephalogram background was categorized as normal, slow, discontinuous, burst-suppressed/burst-attenuated, or suppressed, with the latter two categories combined for statistical analyses. Other electroencephalogram features were described in accordance with American Clinical Neurophysiology Critical Care Electroencephalogram terminology (24), including: reactivity, symmetry, rhythmic and periodic patterns (including generalized periodic discharges, generalized rhythmic delta activity, lateralized rhythmic delta activity, and lateralized periodic discharges), and sleep features; sporadic epileptiform discharges, seizures (focal or generalized), and status epilepticus. Because these findings were documented at the time of patient care, we considered these findings to be blinded to ultimate outcomes.

The primary outcome was dichotomized between “favorable” and “unfavorable” outcomes, with the latter being a composite of mortality and the development of a new disability at the time of hospital discharge (i.e., increase in FSS by ≥ 3 points [21]). Categorical data were expressed as counts, proportions, and percentages, whereas continuous data were expressed as median and interquartile ranges (IQRs). Univariate analyses used chi-square or Fisher exact test (counts < 5) to assess the association between variables and outcome. We next identified potential predictors for the model. All electroencephalogram metrics were candidates for inclusion in the model. GCS score less than or equal to 8 was also included in model building. We also sought to incorporate radiographic findings in our model, used the first head CT at admission. The scans were graded using the CT Rotterdam score, which was developed to predict mortality risk from TBI; the score ranges from 1 to 6 with the higher the value indicating greater risk of mortality (25). A multivariable model was then fitted following a stepwise selection to identify variables with a threshold of inclusion/exclusion of 0.35. Only the variables that met the selection criterion were retained in the final model. A receiver operating characteristics (ROCs) curve for each step was created, in which the predictive model of the final curve was reflected. The area under the curve (AUC) of the final model and the corresponding 95% CI was calculated to summarize the predictive accuracy of predictors. The final model goodness of fit was tested using the Hosmer-Lemeshow test. The sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV) were reported based on the threshold cutoff point that was determined by Youden index.

RESULTS

There were 142 patients eligible for inclusion in our study Figure 1. Demographic information and injury characteristics of the study cohort divided by outcome is provided in Table 1. The median age was 22 months (IQR, 4–79 mo), and 91 of 142 patients (64%) were male. The GCS score was 13–15 (mild TBI) in 62 of 142 patients (44%), 9–12 (moderate TBI) in 20 of 142 (14%), and less than or equal to 8 (severe TBI) in 60 of 142 (42%). Overall, TBI from blunt force trauma or motor vehicle collision occurred in 71 of 142 patients (50%). Of note, 26 of 142 patients (18%) had confirmed abusive head trauma and a further 24 of 142 (17%) had the diagnosis suspected. In the whole cohort, 100 of 142 (70%) did not develop a new disability at hospital discharge, whereas the other 42 of 142 patients (30%) had unfavorable outcome, including eight deaths (i.e., one by cardiac criteria, four by neurologic criteria, and three after withdrawal of technological support) and 34 patients with new functional disability.

Figure 1.

Flowchart of inclusion and exclusion criteria. EEG = electroencephalogram, ICD = International Classification of Diseases, TBI = traumatic brain injury.

TABLE 1.

Patient Demographics and Injury Characteristics

Variable	Favorable Outcome (n = 100)	Unfavorable Outcome (n = 42)	Total (n = 142)
Age, mo	9 (3–59)	45 (15–107)	22 (4–78)
Sex (male)	61 (61%)	30	91 (64%)
Cause of traumatic brain injury
MVC—pedestrian	10 (10%)	9	19 (13%)
MVC—passenger	12 (12%)	10	22 (15%)
Blunt force	18 (18%)	12	30 (21%)
Fall	50 (50%)	5	55 (38%)
Penetrating	0 (0%)	1	1 (< 1%)
Other/unknown	10 (10%)	5	15 (11%)
Abusive head trauma
Confirmed	13 (13%)	13	26 (18%)
Suspected	19 (19%)	5	24 (17%)
Admission Glasgow Coma Scale
13–15	59 (59%)	3	62 (44%)
9–12	15 (15%)	5	20 (14%)
≤ 8	26 (26%)	34	60 (42%)
Admission Functional Status Score	6 (6–6)	6 (6–6)	6 (6–6)
CT Rotterdam score	2 (1–2)	2 (2–3)	2 (1–2)

MVC = motor vehicle collision.

Data are presented as n (%) or median (interquartile range).

The injuries seen on initial head CT in both favorable and unfavorable outcome patients is summarized in the Table S1 (http://links.lww.com/PCC/C570). Overall, 87 of 142 patients (59%) had subdural hemorrhage, 63 of 142 (44%) had subarachnoid hemorrhage, and 16 of 142 (11%) had intraventricular hemorrhage. Table 1 shows that outcome grouping (unfavorable vs. favorable) was associated with worse CT Rotterdam score (p < 0.007). Overall, 44 of 142 patients (31%) underwent a neurosurgical procedure (i.e., hematoma evacuation, decompressive craniectomy, intracranial pressure (ICP) monitor placement), with 33 of 142 (23%) undergoing ICP monitoring. Outcome grouping (unfavorable vs. favorable) was associated with a higher proportion of neurosurgical procedures (26/42 [62%] vs. 18/100 [18%]; absolute difference, 44%; 95% CI, 26.6–58.5%; p < 0.0001). Of the 89 of 142 patients (63%) discharged home, 87 patients (98%) had a favorable outcome. In the 45 of 142 patients (32%) discharged to an inpatient rehabilitation facility, 32 of 45 had an unfavorable outcome. Thirteen patients from the favorable group had a delta FSS of 2 and were sent for inpatient rehabilitation.

Table 2 reviews the cEEG features and their associated outcomes. Overall, 96 of 142 patients (68%) had an abnormal electroencephalogram background. Discontinuous or burst-suppressed/burst-attenuated/suppressed electroencephalogram background activity were associated with lower odds of favorable outcome (odds ratio [OR], 0.18 [95% CI, 0.05–0.63] and 0.10 [95% CI, 0.03–0.38], respectively). Conversely, normal electroencephalogram background and reactivity were associated with greater odds of favorable outcome (OR, 9.81 [95% CI, 2.84–33.86] and OR, 48.55 [95% CI, 14.17–166.08], respectively). Sleep features were also associated with outcome (p < 0.0001). However, we failed to identify an association between outcome grouping and slowing on electroencephalogram, asymmetry, rhythmic and periodic patterns, sporadic epileptiform discharges, and any seizures.

TABLE 2.

Continuous Electroencephalogram Characteristics

Variable	Favorable Outcome (n = 100)	Unfavorable Outcome (n = 42)	p	OR (95% CI)
Background electroencephalogram			< 0.0001	—
Normal	43 (43%)	3	< 0.0001	9.81 (2.84–33.86)
Slowing	50 (50%)	21	1	1.0 (0.49–2.06)
Discontinuous	4 (4%)	8	0.003	0.18 (0.05–0.63)
Suppression or burst suppression/attenuation	3 (3%)	10	0.0003	0.10 (0.03–0.38)
Reactivity (n = 128)	89 (96%)	11	< 0.0001	48.55 (14.19–166.08)
Asymmetry	44 (44%)	22	0.36	0.71 (0.35–1.47)
Rhythmic and periodic patterns	11 (11%)	4	1	1.17 (0.36–3.92)
Sporadic epileptiform discharges	27 (27%)	8	0.32	1.57 (0.65–3.82)
Sleep features (n = 136)			< 0.0001	—
Present	74 (76%)	11		—
Asymmetric	9 (9%)	24		—
Absent	14 (14%)	4		—
Seizures and/or status epilepticus	34 (34%)	12	0.52	1.29 (0.59–2.83)

OR = odds ratio.

Data are presented as n (%). Dashes indicate there are no OR to report.

Next, using stepwise regression, we developed a multivariable model incorporating two injury-associated parameters (i.e., GCS ≤ 8, CT Rotterdam score), the use of sedative infusion, and all captured electroencephalogram metrics Table 3. The following measures met criteria for inclusion in the final model: GCS less than or equal to 8, the presence of sedation, discontinuous electroencephalogram background, reactivity, rhythmic and periodic patterns, and sleep features. The ROC curve for the final model selection is shown in Figure 2 with an AUC of 0.94 (95% CI, 0.90–0.98). In the model, reactivity on cEEG increased odds for favorable outcome, while a GCS less than or equal to 8, the presence of sedative infusion, and the absence of sleep features reduced the odds of favorable outcome. The Hosmer-Lemeshow tested showed a goodness of fit with a p value of 0.27. The PPV of the model was 91.7% with a NPV of 87.1% (Table S2, http://links.lww.com/PCC/C570). On evaluating the diagnostic utility of a positive model result, we start with a population pre-test probability of favorable outcome of 0.70. The post-test probability of favorable outcome rises to 0.91 with a positive model result. Last, the prevalence of favorable outcome in our population is above the prevalence threshold of 0.59 for the model.

TABLE 3.

Parameters for Model Building for Predicting Favorable Outcome

Parameter	p	OR (95% CI)
Glasgow Coma Scale ≤ 8	0.02	0.21 (0.05–0.88)
Sedation	0.0004	0.13 (0.03–0.60)
Background electroencephalogram
Discontinuous	0.19	11.16 (0.85–146.50)
Reactivity	< 0.0001	42.71 (5.37–339.85)
Rhythmic and periodic patterns	0.26	0.34 (0.05–2.34)
Sleep features	0.17
Absent (vs. present)	0.03	0.14 (0.02–0.86)
Asymmetric (vs. present)	0.27	0.34 (0.05–2.30)

OR = odds ratio.

Figure 2.

Receiver operating characteristic (ROC) for stepwise regression. Stepwise selection was used to detect independent variables at the significance level of 0.35 to create the final model. Step 0: addition of intercept. Step 1: incorporated reactivity on electroencephalogram (EEG). Step 2: incorporated sedation infusion. Step 3: Glasgow Coma Scale less than or equal to 8. Step 4: discontinuous EEG. Step 5: sleep features. Step 6: normal EEG. Step 7: rhythmic and periodic patterns.

DISCUSSION

In this single-center retrospective cohort of pediatric TBI patients, we have produced a hypothesis-generating model of favorable outcome using clinical and electroencephalogram data. The model is based on pediatric TBI patients who met the following criteria: 1) managed in our PICU, 2010–2020; 2) admitted to the PICU according to local policy and requirements; 3) no history of a prior seizure disorder; 4) no history of cardiac arrest in association with TBI; and 5) selected for cEEG monitoring that lasted at least 8 hours within 72 hours of PICU admission. In such a population, we have found the pre-test probability of favorable outcome is 70% (or 0.70) and, with a “positive” result in the model, the post-test probability of favorable outcome rises to 91% (or 0.91).

Admission GCS is designed to assess the level of consciousness of a patient with head injury, while universally applied to classify severity of injury, has limited ability to predict outcomes, especially in children (26-29). We considered GCS less than or equal to 8, a validated CT scoring scale for inclusion in the model (25) and the presence of sedative infusions. In our stepwise design, admission GCS less than or equal to 8 met criteria for inclusion, but demonstrated limited ability to predict outcome alone, consistent with previous studies in children (24-27). Similarly, the presence of sedative infusions, an important confounding factor when considering electroencephalogram metrics (30), also met criteria for inclusion. We suspect that the use of sedative infusions is likely a surrogate of injury severity, as sedation is often required in patients requiring invasive mechanical ventilation as well as those treated intracranial hypertension (31). Interestingly, background burst suppression or attenuation did not meet criteria for inclusion, while other metrics of electroencephalogram features did, suggesting that the electroencephalogram metrics included in the model may be more nuanced than a direct correlation with sedative use due to clinical severity. Unexpectedly, the CT Rotterdam score did not meet criteria for inclusion in our final model, despite previously showing a predictive ability in severe pediatric TBI (25). We do note, however, that our patient sample’s Rotterdam score range was small, and its effect may have been outweighed by other aforementioned metrics that did meet our inclusion criteria. Additionally, as prior studies have also shown that initial CT imaging at the time of presentation may be unreliable as a biomarker (9, 10), it is possible likely that, like admission GCS, such early imaging does not provide insight into injury evolution (e.g., increasing size of a hemorrhage, ischemia or edema) which could play a role in determining long-term outcomes. In this regard, cEEG deployed for several hours within the first few days of admission, as brain injury evolves, may provide added predictive ability (32).

Since the initial applications of electroencephalogram for post-TBI neuromonitoring (33, 34), electroencephalogram is commonly used in the ICU as a noninvasive bedside tool to detect seizures (15). In our study, we posit that other electroencephalogram characteristics may lend prognostic insights as has been suggested such as burst suppression, is associated with unfavorable outcome after acute brain injury, especially in patients with hypoxic ischemic encephalopathy; however, abnormalities observed lack specificity (35). Electroencephalogram is more accurate in prediction when used in combination with other features (36, 37). In our study, discontinuity and background suppression/burst attenuation/burst suppression was associated with functional outcome, but no association was found with slowing, a common nonspecific finding on electroencephalogram (36), which was equivalently found in those with favorable and unfavorable outcomes (37). Consistent with earlier studies in brain injured patients (18, 34, 38, 39), the presence of reactivity and sleep on electroencephalogram was predictive of favorable outcomes (20, 31, 35, 39). This data is broadly consistent with a larger study of moderate-severe TBI in adults (18) in which poorer outcomes were associated with absent posterior dominant rhythm, absent sleep transients, predominant delta activity, and a discontinuous background. That study also noted that ictal-interictal continuum patterns correlated well with injury severity but not with functional outcome, which is consistent with our data and others in which seizures failed to show an association with functional outcomes (39). Of note, our study incorporated electroencephalogram findings from the first 72 hours after admission; the relationship between the timing of our noted electroencephalogram features (e.g. 12, 24, 72 hr) and clinical outcomes remains unknown.

While our model produced a high predictive ability for outcome of 0.94 with 95% CI (0.90–0.98) with a post-test probability of 91%, we recognize that this likelihood may fall short in providing bedside providers with the necessary confidence to counsel families of pediatric TBI victims. Furthermore, we also remain cautious in generalizing our findings to patient populations beyond our institution experience, as the enhancement in prognostic ability may depend on an institution patient population and TBI outcomes. Nevertheless, this data is important in that it supports the inclusion of electroencephalogram metrics in future TBI outcome modeling efforts, which can be used in conjunction with clinical examination and brain imaging.

The design of our study obligates it to possess several limitations. As a retrospective chart review from a single site with a relatively small sample size, this study is limited in its ability to include additional factors for modeling, compare our model to other multiple models, and perform validation in a split cohort. The limited numbers also necessitated that we combined disability (FSS ≥ 3) with mortality as a categorical “unfavorable outcome,” a larger study would also be better suited to differentiate electroencephalogram metrics that may contribute to either outcome, individually. Additionally, as a single-center study, it is subject to institutional practices that are not pathway/algorithm defined, including the initial decision for admission to the PICU vs. other inpatient floors, the decision to obtain cEEG, and the decision to proceed to neurosurgical intervention. However, because not every TBI patient admitted to the PICU undergoes cEEG, our dataset likely reflects a bias toward obtaining cEEG preferentially on those clinically expected to have worse outcomes. The benefit of a follow-up multicenter study will be able to address these shortcomings.

As researchers continue to refine prediction models for pediatric TBI outcomes, we expect that the application of machine learning to evaluate larger datasets from multiple centers will incorporate more sophisticated cEEG features, both qualitative and quantitative, in prediction algorithms. These technologies will be capable extracting nuances from cEEG records, brain imaging, and, importantly, long-term morbidities such as specific neurocognitive outcomes gleaned from formal neurologists’ assessments. We anticipate a future in which bedside providers can receive continuous feedback as evolving clinical data is provided to more accurately discriminate between favorable and unfavorable outcomes, ultimately informing the care we provide to our patients and their families. The inclusion of such a wide net of data has already proved useful in developing algorithms capable of predicting the onset of critical illness in pediatric patients, which can be further retrained for single-site application (40).

Our study showed specific cEEG findings observed acutely after injury, in combination with other clinical characteristics, have the potential for predicting short-term functional outcomes after pediatric TBI. Future studies adopting a prospective approach across multiple centers will more definitively confirm these associations, with a goal of providing bedside clinicians with advanced capability to accurately discriminate between favorable and unfavorable outcomes.

Supplementary Material

Supplemental digital content is available for this article. Direct URL citations appear in the printed text and are provided in the HTML and PDF versions of this article on the journal’s website (http://journals.lww.com/pccmjournal).

RESEARCH IN CONTEXT.

• Pediatric traumatic brain injury (TBI) is one of the leading causes of death and disability in children, with increasing recognition of its contribution to persistent neurocognitive impairments in survivors and their impacts on families and society.

• Neuroprognostication of functional outcomes in pediatric TBI is challenging due to the heterogeneous injury mechanism and developmental neuroplasticity.

• Data from continuous electroencephalography (cEEG), often used to monitor seizures after TBI, may provide additional insight that is valuable for neuroprognostication following neurotrauma.

WHAT THIS STUDY MEANS.

• cEEG features, such as background, reactivity, rhythmic and periodic patterns, and sleep features are associated with clinical outcomes following TBI.

• Although abnormal cEEG features are associated with unfavorable outcome, they lack specificity for neuroprognostication when used in isolation.

• A combination of early qualitative cEEG features and clinical characteristics such as Glasgow Coma Scale in pediatric patients with TBI have been used in a hypothesis-generating model about short-term functional outcomes. Validation is now required in a different population.